SO-100 (SO-ARM100) · published weights

SmolVLA base (pretrained VLA to fine-tune on your own SO-100 dataset)

Record

Robot: SO-100 (SO-ARM100) therobotstudio/so-100
Framework: LeRobot (SmolVLA) · LeRobot
Policy license: not stated by the package · robot software repos: Apache-2.0
File: model.safetensors
Source: https://huggingface.co/lerobot/smolvla_base
Note: 450M. Per the SmolVLA paper (arXiv 2506.01844), pretrained on community LeRobot datasets that are mostly SO-100 recordings. The LeRobot docs demo fine-tuning on lerobot/svla_so100_pickplace (50 SO-100 episodes, 5 cube positions). Base model, not zero-shot. Apache-2.0.

Not in the tab yet

To replay this policy in the browser we need an ONNX export of the named checkpoint, the observation and action contract copied from the authors' own sim2sim or deploy script, and a MuJoCo model of the robot that the contract matches. Then it must walk 60 s in our runtime without a fall. Until then the source link is the way in.

Other SO-100 (SO-ARM100) policies · 1

SO-100/101 language-conditioned manipulation (VLA, zero-shot or fine-tune)LeRobot (MolmoAct2)